Closed set and open set Speaker Identification using amplitude distribution of different Transforms

Abstract

In this paper, closed set and open set Speaker Identification has been performed on two different databases. Feature extraction for the Identification has been done by using the amplitude distribution of four different Transforms i.e. DFT, DHT, DCT and DST. Two similarity measures i.e. Euclidean Distance (ED) and Manhattan Distance (MD) have been used for matching. The performance has been compared with respect to the best value of each Transform for following parameters: length of speech sample, similarity measure score, size of feature vector, FAR/FRR performance and data acquisition system. Amongst the transforms the best result is given by DFT at 99.06% for feature vector of size 32. Amongst similarity measures Manhattan distance outnumbers the Euclidean distance by 54 to 15, considering the results of all three lengths of speech. The best GAR is 90.65% with a threshold of 94.11% for DFT with MD as a similarity measure.